Artificial Intelligence has become the defining investment theme of this decade. Yet history suggests that transformative technologies rarely create value in a straight line.

In this latest paper, Dr Sandy Nairn examines today's AI boom through the lens of previous technological revolutions, from canals and railways to the internet. While he remains firmly positive on the long-term impact of AI, he argues that investors should distinguish between the transformative potential of the technology itself and the valuations currently being attached to many of its leading companies.

For Goodhart's investment team, the question is not simply whether AI will change the world, but where future returns are likely to emerge as the technology evolves, capital cycles mature and today's winners are tested by competition and commercial reality.

Bubbling Up

The arrival of trillion-dollar AI IPOs, for the moment, do not appear to have destabilised markets.  Granted for SpaceX, the first arrival, the initial free float was pitched at a level to avoid indigestion, and its post listing price was assisted by technical factors relating to its forthcoming inclusion in the NASDAQ index. The technical underpinning no doubt served to further increase initial investor appetite, but as more recent moves highlight, it has a limited shelf-life.  How sustainable and generalisable the original IPO valuation may be is a different question.  The valuations and the magnitude of the coming IPOs, together with the valuations of existing AI-related companies have intensified the ongoing debate and narrative about bubble valuations.

A persuasive overheating case can certainly be made, both in contemporary terms and with reference to many historic markers.  Negative commentary has increased in intensity as share prices have risen and new financings have been announced.  This is not unusual.  Warnings often precede market declines, but before any falls the market can prove particularly stubborn and resistant to any warnings as FOMO becomes a key driver.  Old saws such as, ‘markets are a voting machine in the short-run and a weighing machine in the long-run’ proliferate with a constant tug of war between the two uppermost in investor’s minds.  It is a cycle that has a powerful historic pedigree.

This particular technology iteration rests upon a capex boom of historic proportions as the infrastructure necessary for AI is built out at breakneck pace.  The speed of deployment has produced many supplier bottlenecks, allowing profit margins to rise with revenues and reach levels significantly above what would historically been perceived as normal.

With many generative AI companies still loss making and lacking sufficient revenues, the AI infrastructure build-out has been financed in anticipation of future revenue flows. A combination of private equity, a complex network of vendor financing, and third-party debt funding has been required.  Arguably this has been sustained by sharply rising valuations of the LLM foundation model companies.  This in turn has been underpinned by rapid user growth, some of which has been revenue generating.  In the TMT bubble ‘eyeballs’ became the proxy metric for future revenue streams.  When they initially failed to translate to actual payments in anticipated volumes, the absence of cash-flow led to the vicious bear market that followed.  There are parallels with current circumstances.  User growth and token use have followed an exponential path underpinning the excitement for future revenues.  The principal question is whether they will translate to revenues before investors get twitchy, or whether there will be an ‘eyeball’ repeat.

It is important to separate out two different questions.  The first relates to the overall impact of AI as a technology, the second relates more specifically to the LLM technology cycle as manifested in the recent and forthcoming IPOs. It is important to stress that two different views can comfortably co-exist.

Long-term technology: the cognitive revolution

There is an equivalency to the industrialisation age in what we are now witnessing, in the sense that for the first time we have the coming together of same three ingredients: power (compute), raw materials (data), and machines (algorithms). The parallel I would draw is impact of James Watt separating out the condenser in the steam engine thus tripling its efficiency.  This set the stage for multiple technology power cycles which ignited and sustained the 150-year period of industrialisation.

In this sense, it can be argued that we are only at the beginning of what could loosely be termed a cognitive revolution, one that will unfold over the coming decades. The evidence is persuasive that the technology will prove both enduring and transformative.  It is already displacing existing methodologies and will continue to do so as it develops further and disseminates through to an ever-wider range of uses.  The early losers in this game are already apparent, even if some stock market reactions have been somewhat indiscriminate. More losers will follow.

Each technology iteration has its own cycle, and hence it is important to be reminded that the bubble question is specific to that cycle and also to a particular point in time.  The danger for bullish investors is that a technological change as profound as AI is taken to imply blue skies not only for AI itself, but for all AI-related companies.  This is not the lesson of history.  The slide below is drawn from the final chapter of the coming book. To reiterate the main points: we are at the early stages of a technological evolution, the impact of which will be profound and will persist for many decades to come.  Within this evolutionary process there will be a number of technology capital cycles.

Figure 1 The Cognitive revolution

For the global economy, the deployment of AI will ultimately prove to be incredibly positive, but just as the industrial revolution included both multiple bubbles and crashes for investors, so too will the cognitive revolution.  So where do we sit in the current LLM generative AI cycle?

Field of Dreams – ‘if you build it, he will come’

This article draws both from the final summary chapter of the forthcoming book and the research contained within Engines that Move Markets. The Engines book was written in 1999-2000 (with a second edition in 2018) as the TMT bubble was inflating.  It looked at historic technological advances, seeking to identify any repeating patterns and hence draw out any lessons could usefully be learned.  The period it covered was from the building of canals in the early 1800s through to the TMT period.  It therefore encompassed the shift from an agrarian subsistence society through the 200 years of industrialisation. 

The stages of historic technology cycles it identified are outlined below in Figure 2.  The cycle begins with the initial concept/feasibility and ends with ultimate financial success or failure. As substantial capital is required to progress from initial concept to prototype and ultimately  to commercial deployment,   promotion and hype are often unavoidable  features of of the jounrey.  This cycle is no different.

The term AI covers a wide waterfront and just for clarity, the current bubble question seems to specifically target today’s LLM phenomenon spanning LLM foundation models, chatbots and agentic applications together with the enabling infrastructure. There are many other areas of AI which are not yet nearing the public markets. 

Although technology cycles have all followed the same general pattern, they have not been identical.  In this case, the AI cycle differs slightly in that the surge in IPOs has been deferred, with listed market excitement instead largely devoted to compute power and infrastructure companies. The IPO element has been delayed as capital-rich private backers sought to garner as much of the valuation uplift as possible. The companies have now reached the stage where the volume of funding required cannot be satisfied from private capital alone.  Public capital is now required to fulfil ongoing capital needs to cover the continued infrastructure build-out until the revenue and profit gap closes.  For the initial private capital investors, there is also the imperative to provide a potential exit and liquidity for existing private capital investors. Rather than a constant stream of companies coming to market to take advantage of public funding, this is more akin to dam bursting, with a smaller number of companies arriving at truly extraordinary scale.  Between SpaceX, Anthropic and OpenAI, the cumulative market capitalisation at listing is likely to exceed $4 trillion. 

So where are we now?

We are firmly in Stage 3: the funding and commercial viability stage.  Capital raises are critical to bridging the funding gap between now and eventual profitability.  At this stage IPO companies need to publicly disclose financial forecasts as part of the listing process.  The pulling back of the curtain both exposes them to comparisons with existing listed companies and sets criteria against which they will be judged.

Figure 2 The technology cycle

The next stage is where the rubber hits the road; where the markets discriminate, where valuations reset and the crowd gets severely thinned out. In this cycle there is the complication of cross linkages and aspects of vendor financing aspects which are likely to introduce systematic risk through the financial transmission mechanisms.  This will be addressed more specifically in a later piece, but for the moment it is sufficient to use historic examples as reference points.

Noteworthy recent examples

What follows may appear to state the obvious but nevertheless it bears repeating.  Whilst incredibly successful companies emerge from the technology cycle, they tend to do so from a crowded field where the most common outcome is failure.  Rather than go through all the financials this is illustrated by reference to the companies which populated the field, most of whom no longer now exist.  The two examples are the emergence of GPUs and Search.  One critical point is that whilst each participant had a possibility of individual success, there had to be a high failure rate since aggregate potential revenues could only satisfy a small number of the competitors.

Figure 3: GPUs - E pluribus unum

The development of GPUs followed the expansion of the gaming industry as PCs became gaming platforms and specialist gaming consols sought to capture the home market.  Many companies attempted to capture this growth, but eventually only a handful remained around long enough to benefit from AI as the new explosive use case.  For many years it was not at all obvious that it would be Nvidia who would remain to survey the wreckage.

Figure 4:  Search is not hard to find

Exactly the same unfolded with Search.  Early leaders were overtaken by Google who then came to dominate the market, benefitting from its natural monopolistic characteristics.

As financial reality raises its head during the rationalisation and refinancing phase, sentiment rapidly reverses.  Capital becomes much more descrimminating and many potentially viable companies suffer, leadng to either their exit or rationalisation.

Figure 5: Eventually sentiment turns

Positive sentiment cannot be sustained in period of corporate failure and this often coincides with a rising cost of capital.  It is easy to forget how quickly positive certainties can be reappraised into tales of doom and gloom.

So, what do investors need to consider today?

In the current environment we have to add the complication of how much the private backers have left on the table for public investors.  Some of the most astute investors in the world, investors who are party to all the prior internal discussions and changing financial projections, have decided that now is the time to IPO.  In part, this reflects the need for these companies to raise more capital than private investors alone can supply. In part it may simply be that existing investors have decided that, on a risk-reward basis, now is a good time to lighten their exposure.

Users and revenues may have been rising sharply but so too have valuations, and ultimate profitability remains a distant goal for many. It may be a new technology, but the competition is intense and well-funded.  This competition will come not just for the companies who are coming to the market, but existing well-funded astute competitors.  Google and Meta for example have each either raised, or are expected to raise, circa $80bn additional capital to supplement self-funded AI development costs.  When cash-flow positive companies raise such eyewatering amounts it emphasises the importance they attach to AI for their future survival and development. It also reminds us of the capital intensity of AI.

Further, the importance of AI extends beyond the individual enterprise, it is also a geopolitical hot button.  It is to be expected that more than just the US and China will perceive the dangers of being dependent upon other power blocs for access.  This suggests government involvement and capital will continue to flow for some time.  This can only serve to extend the time and cost of the rationalisation phase.

Figure 5: Do not need to search for LLMs either

Much of the debate naturally focusses on the leaders such as OpenAI and Anthropic but as Figure 5 illustrates, many companies are involved in developing LLMs and ancillary services.  There are differences in their scale and their level of support but suffice to say, it is not an uncrowded field.

With numerous well-funded participants, the scene is set for an extended period of bleeding as the combatants battle for client revenues.  Whilst arguments can be made in support of the current valuation (even if a bit stretched) of any individual LLM participant, no sustainable argument can be made for the collective revenues that would be required for all to succeed.  Investors therefore need not only to judge the potential profitability of the technology and its timing, but also who is going to emerge triumphant and whether the risk/reward profile on current valuations is favourable.

Part of the selection process depends upon who is likely to prosper in both the retail and enterprise segments.  On the enterprise side, the current focal point is software development and coding.  Whilst it is straightforward to identify the current leader, Anthropic, the sustainability of that leadership position is another question.  To try and get a feel for predictability, over the last three years I have spoken to AI technology specialists about their use of the leading LLM providers and how they rank in terms of product quality and usability.  Suffice to say, those rankings have changed dramatically from year to year.  This suggests that the jury remains out on who will win the race in generative and agentic AI, irrespective of who appears to be leading today. What these specialists also highlighted is that their choice of provider is largely a straight-forward value-for-money decision, and they are prepared to switch providers as appropriate.  In other words, the race may be a perennial one, but with fewer competitors and persistent pricing pressure.

On top of company-specific risk there are also potential systematic risks.  Whilst vast sums have been raised, the interlocking of funding carries with it a potentially dangerous transmission mechanism if defaults or failure were to emerge.  This will be discussed in a future paper.

The reality is that there is unlikely to be significant shareholder upside from current levels, irrespective of who eventually emerges as the winners/survivors.  For the losing companies the downside will range from savage to extinction. Patience will be key. It is also important to bear in mind that AI extends beyond the big LLM companies, and there will also be a plethora of new specialist companies coming to the market - all with exciting prospects. This should not be forgotten when sentiment turns negative, focussing on these areas now is important as it lays the groundwork which will later pay dividends.

The increasing narrative over bubble valuations matches what has been witnessed in previous technology cycles.  The debate is not so much about whether excess valuations exist, as much as what should be expected when cash-flow becomes a binding constraint and capital destruction begins.  The winners will then be clear and likely exist at more appropriate risk/reward valuations.  However, it is entirely possible that even while this is unfolding other AI segments will be searching for capital, but in a much less forgiving environment – which will present highly attractive investment opportunities. In other words, whilst it may be fair to say that there is at least one investment pot on the boil, there are other segments where the temperature is considerably cooler.

Against what might appear a rapidly growing narrative of negativity it must be remembered that we remain in the early stages of the emergence of a truly transformative technology.  Some share prices may be hyped, but the technology is not.  To repeat the comment made earlier, there is no reason why one cannot subscribe to both views simultaneously.